TW201911812A - Obstacle positioning system and maintenance and operation method of video streaming service for the maintenance and operation unit to select the required maintenance information according to the timeliness or the accuracy requirement by inputting plural characteristic parameter sets - Google Patents

Abstract

The invention discloses an obstacle positioning system and maintenance and operation method of a video streaming service. When there is a new obstacle event area positioning problem to be processed, the model which is trained and optimized in advance can be used to automatically generate the binary obstacle area positioning information. Also, by inputting plural characteristic parameter sets, the maintenance and operation unit can select the required maintenance information according to the timeliness or the accuracy requirement to perform maintenance and repair of the equipment. In addition, based on the comparison between the current and past statistical values of the obstacle areas, the maintenance priority plan decision information is outputted to achieve the maximum benefit to the maintenance resource allocation.

Description

 Obstacle positioning system for image streaming service and maintenance method  

The invention belongs to an obstacle positioning system and a maintenance method for video streaming services, in particular to a plurality of Internet service providers (ISPs) that utilize equipment quality measurement, user declaration, repair history record, and manual test. The data source of the industry is combined to establish a classification model of the number of high-dimensional parameters.

With the increasing popularity of Internet broadband services and the large-scale increase in users, the network structure and components of ISP operators have also grown significantly, which has made the difficulty of handling obstacles increase.

On the contrary, in order to facilitate the management and deployment of maintenance manpower, many ISPs are moving towards the direction of flat maintenance system. Therefore, it is a much-needed organizational requirement for ISP operators to quickly determine obstacle areas and call for appropriate human and material resources.

The application of machine learning is an area that is gradually being valued by ISPs. In the early days, it was widely used in the field of e-commerce as a recommendation engine and advertising. In recent years, such as medical science, weather forecasting and many other fields are also popular applications.

With the increasing popularity of applications with huge amounts of data, the way ISP (Internet Service Provider) operators monitor and manage end-to-end network quality must enter the global domain by means of previous sampling tests or purchase of monitoring modules. The stage of management and monitoring.

Most of the current obstacle reporting systems still need to be manually operated by the customer service staff, in order to ask and record the detailed cause of the fault. However, the relative disadvantage is that when manually selecting the reason code, sometimes it is inevitable that mis-selection or inaccuracy will occur (such as the reason for the selection is unknown or otherwise). At this point, the original description of the obstacles can play the role of auxiliary classification, and make up for the lack of information about the reason code of the application reason input by the customer service staff.

In general, ISP's network services use a mesh structure. However, video streaming services usually use a tree structure to reduce the interval between the amount of data transmitted and the feedback of feedback.

Due to this tree structure, for each user terminal device, its data transmission path is fixed from the image server to the terminal device itself. When a user experiences a problem with the video streaming service, the obstacle point will roughly fall on the related device of the path. Therefore, for the analysis of obstacle points in a single report, linear regression models have been used in the past to analyze equipment and related data on fixed paths, and to establish a model of obstacles that are initially credible.

However, in order to achieve higher accuracy of obstacle location and priority decision requirements, after the source of data sources has increased significantly, the general linear regression model is insufficient to meet the needs of ISPs for processing, and nonlinear models capable of processing high dimensional parameters must be used. To work.

In view of the shortcomings derived from the above-mentioned conventional methods, the inventors of the present invention have improved and innovated, and after years of painstaking research, they finally succeeded in developing and developing the obstacle positioning system and maintenance method of the video streaming service.

In order to achieve the above object, the present invention provides an obstacle positioning system and a maintenance method for video streaming services, which are characterized by high-dimensional quantity eigenvalue extraction and model training by various types of integrated network management information established by ISP operators. In order to find the best model after optimization, and use this model to generate forecasting information for the future obstacles customers based on the data obtained at the time. It can also provide a priority decision-making scheme for adjusting the maintenance of video streaming services to improve the efficiency of the use of limited maintenance resources.

The ISP industry itself establishes a variety of high-dimensional eigenvalue combination classification models based on equipment and line product quality measurement, user declaration, repair history record, and manual test. In the future, whenever a new obstacle event occurs, the number of high eigenvalues trained beforehand can be applied to the nonlinear regression model, and the model can be selected according to the current combination of eigenvalue data and the timeliness or correctness of the maintainer. Estimate the information of the obstacle area, process it, and produce a plan to help adjust the maintenance priority decision. The defined obstacle area is divided into the client (including the user's home loop and terminal equipment) and belongs to the ISP. The end equipment (including the equipment-to-equipment loop) is two categories, which means that the obstacle can be located in which part of the obstacle is analyzed by the method of the present invention, and the ISP is applied according to the maintenance organization assigned maintenance mode.

An obstacle positioning system for video streaming services, comprising: a data source module, which is a plural information source required for collecting and determining obstacle positioning, and further comprising a service quality management unit, which is a service quality data including an application layer, The image quality level and the numerical quality index of each image terminal device; the obstacle report management unit is the information related to the obstacle application including the image service, and is the content of the report, the description text and the test code after the artificial pre-test; the loop quality diagnosis The management unit is a loop quality test data including the physical layer and a non-standard user loop construction method record. It is a special method for estimating the electrical characteristics of the line and the recording of the ISP's switching office equipment closest to the user end, which greatly affects the user's use distance. Records such as line coupling (bundling) and optical copper hybrid (G.fast); broadband network monitoring unit, is the information of the brand model data, equipment alarm code and alarm code and content of each node equipment of the ISP.

The feature value extraction module extracts various types of information required for positioning the obstacle according to different source system characteristics to form an input feature parameter group of the subsequent machine learning analysis module, and processes each unit in the extracted data source module. The feature value is another service quality management feature value extraction unit, which is to obtain the binary flag value of the high-definition user of 4K or more in recent days, the video stream service quality index, and the video stream service appeal probability; The management feature value extraction unit is to obtain the video stream service request reason code, the report description text note, the manual diagnosis pre-test code, and the obstacle reason code input after the preliminary manual test for the professional diagnostic personnel; the loop quality diagnosis management feature The value extraction unit is to obtain the Digital Subscriber Line Access Multiplexer (DSLAM) brand model, the DSLAM firmware version, the voice and audio band attenuation value, the uplink SNR (Signal to Noise Ratio), and the downlink SNR. , the quality monitoring value of the user and the ISP end periodically, whether to use the special method Such as line coupling (bundling) and optical copper hybrid (G.fast) method binary flag value; broadband network monitoring feature value extraction unit, is to obtain the image of the video streamer box or home multi-function gateway, The downstream rate of the video streamer box or the home multi-function gateway, the type of the central office equipment, the quantized value of the alarm value of the central office equipment, the term frequency of the alarm type, and the alarm severity index value.

The machine learning training and implementation module is to accept the characteristic values of the feature value extraction module output, further perform data preprocessing, train with machine learning and obtain optimized model parameters, and further include a training unit and An implementation unit, wherein the training unit further comprises a training target establishing unit, which is a criterion for establishing a prediction model, as a calculation loss function in the training model process and a benchmark when optimizing; the category and the missing value pre-processing training unit are The eigenvalues of the training data are preprocessed, and include expanding the categorical eigenvalues into binary indicator eigenvalues, and, if the numerical eigenvalues have missing values, replacing them with averaging values, and The missing feature value adds a binary missing indication feature value; the text note obstacle analysis training unit is a verbatim note content for each obstacle pending event in the training data, using the automatic word breaking tool and Logis regression analysis, first calculate the text description of the barrier-related word frequency combination is the client or ISP end barrier Obstacle probability, and take the probability as one of the characteristic values of the subsequent model input; the high-dimensional eigenvalue multi-combination establishes the training unit, which is to make one or a plurality of high-dimensional feature value sets for each obstacle waiting event of the training data; The module for optimizing the classification model of obstacles is to use the non-linear gradient decision tree (GBDT) as the main estimation model, and input multiple combinations of high-dimensional eigenvalues to establish the high output of the training unit. After the dimension feature values are combined, the optimal model parameters of the training data are found through the optimization process of minimizing the loss function, and each newly added obstacle interval data is estimated for actual application.

The implementation unit of the machine learning training and implementation module is responsible for the actual real-time data processing estimation, and the pre-processing unit for the category and the missing value is to preprocess the characteristic values of the actual data to be estimated and the characteristic values. And including expanding the categorical feature value into a binary indicator, and if the numerical eigenvalue has a missing value, replacing it with an average value and adding a second to the partially missing feature value The missing element indicates the feature value; the text note obstacle analysis unit is a verbatim note content according to the text description of each of the actual pending data and the feature value to be processed, using the automatic word breaking tool and the return of the Rogers Analysis, first calculate the word description of the barrier-related word frequency combination is the probability of belonging to the client or ISP-side obstacles, and take the probability as one of the characteristic values of the subsequent model input; the high-dimensional eigenvalue multi-combination unit is for the actual Estimating data and eigenvalues for each obstacle pending event, making one or more sets of high-dimensional eigenvalues; The estimated output unit uses the best optimized GBDT model parameters after monthly update training to calculate the judgment value of the obstacle area probability for each new case to be judged.

The maintenance operation information output module is used to evaluate different input feature value sets and optimize the model error rate after the training phase. The output integration information and maintenance priority plan are used by the maintenance and management personnel respectively.

The maintenance operation information output module further includes a model performance indicator establishing unit, which is a benchmark calculation model for evaluating the quality of the prediction model based on an evaluation of the weighted error rate; the model error rate calculation unit is to utilize the model performance indicator The weighted error rate calculation formula of the unit is established, and the model prediction error rate corresponding to each type of feature value set after the optimization in the machine learning training phase is calculated; the maintenance information output unit is for integrating the output of the customer data to be processed, The obstacle interval judgment result and the reference model error rate are used to provide the maintenance personnel to use the corresponding inspection suggestion information according to the priority of timeliness or correctness; the maintenance priority plan output unit is based on the current month to the past month. Compared with the statistical average of the barriers within the organization, the output of the maintenance priority decision scheme in the organization and the ISP is used to enable priority treatment and repair in more recent fault areas, so as to achieve maintenance resources. Maximum benefit.

A method for locating and locating an image stream service includes: step 1: the data source module extracts the high-dimensional types of estimated feature values by the feature value extraction module; and the second step, through machine learning training and real For module processing, the optimized prediction model is trained with the training data, and the corresponding client and ISP terminal under multi-type eigenvalues are estimated based on the eigenvalue data of the actual customer complaint case. Estimating the probability of obstacles; Step 3: Finally, the maintenance operation information output module is responsible for generating the operation mode selection and priority decision information; wherein the process of machine learning training and implementation module processing in step 2 includes: step one, whether The training model is generated. If yes, the first simulation training is performed first, and the training target is established. If not, the category and the missing value are processed; and after the training target is established, the category and the missing value are entered. Processing training; Step 3, text note obstacle analysis training; Step 4, high-dimensional feature value multiple combination to establish training; Step 5, optimizing obstacle points The classification model is established and returned to the category and the missing value pre-processing; step six, text note obstacle analysis; step seven, high-dimensional eigenvalue multiple combination establishment; step eight, obstacle point classification estimated output; step nine, judge whether to calculate The next user, if yes, returns to whether the training model is generated, and if it is no.

The process of the operation mode of the third step includes: step one, evaluation of the performance index of the design model, and no change after the first design is completed; step 2, calculating the model error rate according to the reserved test data of the latest training stage; 3. The output shall be provided to the maintenance personnel for the transportation information; in step 4, the output shall be provided to the management personnel for the maintenance priority plan information; in step 5, it shall be judged whether there is a next user to be estimated and calculated, and if so, Then, return to step 2, calculate the model error rate according to the reserved test data of the latest training phase, and if not, the process ends.

The invention provides an obstacle positioning system and a transportation method for video streaming services, which have the following advantages when compared with other conventional technologies:

1. It can handle the establishment of high-dimensional video stream service feature value prediction model.

2. It can quickly classify and classify audio and video streaming service obstacle areas.

3. Allow maintainers to choose the recommended maintenance area based on timeliness or correctness.

4. It can provide the priority decision scheme for dynamically adjusting the maintenance of video streaming services, and achieve the maximum benefit of using limited maintenance resources.

110‧‧‧Source Module

111‧‧‧Service Quality Management Unit

112‧‧‧ Obstacle Report Management Unit

113‧‧‧Circuit Quality Diagnostic Management Unit

114‧‧‧Broadband network monitoring unit

120‧‧‧Characteristic value extraction module

121‧‧‧Service Quality Management Feature Value Extraction Unit

122‧‧‧ Obstacle complaint management feature value extraction unit

123‧‧‧Circuit quality diagnostic management feature value extraction unit

124‧‧‧Broadband network monitoring feature value extraction unit

130‧‧‧Machine Learning Training and Implementation Module

131‧‧‧ training unit

1311‧‧‧ Training target establishment unit

1312‧‧‧Class and Missing Value Pre-Processing Training Unit

1313‧‧‧Text note obstacle analysis training unit

1314‧‧‧High-dimensional eigenvalue multi-combination to establish training unit

1315‧‧‧Optimized obstacle point classification model building module

132‧‧‧ Implementation unit

1321‧‧‧Class and Missing Value Pre-Processing Unit

1322‧‧‧Text note obstacle analysis unit

1323‧‧‧High-dimensional eigenvalue multi-combination building unit

1324‧‧‧Identification of obstacle points

140‧‧‧Maintenance Operational Information Output Module

141‧‧‧Model performance indicator building unit

142‧‧‧Model Error Rate Calculation Unit

143‧‧‧Weiyun Information Output Unit

144‧‧‧Maintenance priority programme output unit

S310~S330‧‧‧Process

S410~S440‧‧‧ Machine learning training and implementation module processing flow

S510~S550‧‧‧ Maintenance operation mode

Please refer to the detailed description of the present invention and the accompanying drawings, which can further understand the technical content of the present invention and the purpose of the present invention. The related drawings are: FIG. 1 is an obstacle positioning system and a method for maintaining the video streaming service of the present invention. FIG. 2 is a schematic diagram of a machine learning training and implementation module structure of an obstacle positioning system and a maintenance method for an image streaming service according to the present invention; FIG. 3 is an obstacle positioning system and a maintenance method for an image streaming service according to the present invention; FIG. 4 is a flowchart of an obstacle positioning system and a method for maintaining an image streaming service according to the present invention; FIG. 5 is a flowchart of an obstacle positioning system and a method for maintaining a video streaming service according to the present invention; Flow chart of learning training and implementation module processing; FIG. 6 is a flow chart of the obstacle positioning system and the maintenance method of the image streaming service of the present invention.

The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

The present invention is further described with reference to the accompanying drawings. Referring to FIG. 1 , an architecture diagram of an obstacle positioning system and a maintenance method for an image streaming service includes a data source module 110, which is a collection and judgment obstacle positioning station. The multi-type information source required further includes a service quality management unit 111, which is a service quality data including an application layer, and is a picture quality level (for example, high-resolution HD, 4K or 8K) and a numerical quality indicator of each image terminal device. The obstacle report management unit 112 is related to the obstacle report related information of the image service, and is the content of the report reason, the description description text and the test code after the manual pre-test; the loop quality diagnosis management unit 113 is a loop quality test including the physical layer. Data and non-standard user loop construction method records, which are estimates of the electrical characteristics of the line and the records of the exchanges of the ISPs that are closest to the user end, and the special method records that greatly affect the user's distance, such as the line coupling using Ohm's law (bundling) The construction method can greatly extend the use distance, and the optical copper mixture proposed by the International Telecommunication Union (ITU) The (G.fast) method can increase the rate but greatly limit the user's distance; the broadband network monitoring unit 114 is information including the brand model data, the device alarm code, and the alarm code and content of each node device of the ISP.

The feature value extraction module 120 extracts various types of information required for the obstacle location according to different source system characteristics to form an input feature parameter group of the subsequent machine learning analysis module, and processes each of the extracted data source modules. The feature value of the unit is another service quality management feature value extraction unit 121, which is to obtain a high-definition user flag value of 4K or more in recent days (for example, if the flag value is True, the user has a picture quality of 4K or more, Conversely, it is False), the video stream service quality indicator, and the video stream service appeal probability; the obstacle report management feature value extracting unit 122 is to obtain the video stream service report reason code, the report description text note, and the manual diagnosis pre-test code. After the preliminary manual test for the professional diagnostician, the entered obstacle reason code; the loop quality diagnostic management feature value extracting unit 123 is to obtain the digital subscriber line access multiplexer (DSLAM) brand model. , DSLAM firmware version, voice and audio band attenuation value, uplink SNR (Signal to Noise Ratio), under SNR, user end and the quality monitoring value of the ISP end device closest to the user end, whether to use special methods such as line coupling (bundling) and optical copper hybrid (G.fast) method binary flag value, special method The use will greatly affect the correlation between the obstacle determination and the distance, so it needs to be recorded in advance to be used in the subsequent training model. The broadband network monitoring feature value extraction unit 124 is to obtain the video streamer box or the home multi-function gateway. The model, the downlink rate of the video streamer box or the home multi-function gateway, the type of the central office equipment, the quantized value of the alarm value of the central office equipment, the term frequency of the alarm type, and the alarm severity index value.

In summary, the data source module 110 is used as an analysis data source, and after the feature value extraction module 120 extracts each type of estimated feature value, the machine learning training and implementation module 130 is used to train the training data set first. The optimized estimation model is used to provide the corresponding estimation probability under the multi-type eigenvalues according to the actual data to be tested in the subsequent implementation, and finally the maintenance operation information output module 140 is responsible for generating the maintenance operation. And the integration information needed for priority decision making.

The collection of data sources comes from the collection of various types of information needed to locate obstacles. The management and diagnostic systems that are used when providing image services to ISPs are only logically distinguished. The entities can be built in the same Within the host group or system.

The extraction of the feature values is to select various types of information required for the obstacles, and extract them according to their different source system characteristics to form an input feature parameter group of the subsequent machine learning analysis module. The extraction process includes an independent discriminant screening process. The input data source does not have a significant positive correlation between the correlation tests, or the null hypothesis of the verification method is not discriminant between the samples. After the verification, the p value is less than the significant level of 0.05. Those who reject the null hypothesis are included in the set of feature values.

Please refer to FIG. 2, which is a schematic diagram of a machine learning training and implementation module of an obstacle positioning system and a maintenance method for an image streaming service according to the present invention. The machine learning training and implementation module 130 is configured to accept feature value extraction. The eigenvalues generated by the module 120 are further processed by the machine learning to obtain the optimized model parameters, and further include a training unit 131 and an implementation unit 132, wherein the training unit 131 further includes training The target establishing unit 1311 is a criterion for establishing the prediction model, and is used as a reference for calculating the loss function and the optimization time in the training model process; the category and missing value pre-processing training unit 1312 is for preprocessing the eigenvalues of the training data. And including expanding the categorical feature value into a binary indicator, and if the numerical eigenvalue has a missing value, replacing it with an average value and adding a second to the partially missing feature value The missing element indicates the feature value; the text note obstacle analysis training unit 1313 is for each event to be processed in the training data, according to the text description The verbatim note content, using the automatic word-breaking tool and the Logis regression analysis, first calculate the obstacle-related word frequency combination of the text description is the probability of the client or the ISP-side obstacle, and take the probability as the eigenvalue of the subsequent model input. One of the high-dimensional eigenvalue multi-combination establishing training unit 1314 is to make one or a plurality of high-dimensional feature value sets for each obstacle waiting event of the training data; the optimization obstacle point classification model building module 1315 is utilized The non-linear gradient decision tree (GBDT) is the main estimation model. After inputting high-dimensional eigenvalues and multiple combinations to establish the combination of high-dimensional eigenvalues of the training unit, the optimization of the loss function is minimized. The process finds the best model parameters of the training data, and estimates each newly added obstacle interval data for actual application.

The implementation unit 132 of the machine learning training and implementation module 130 is responsible for performing actual real-time data processing estimation, and further includes a category and missing value pre-processing unit 1321 for pre-predicting the actual estimated values of the data and the feature values. Processing, and including expanding the categorical feature value into a binary indicator, and, if the numerical feature value has a missing value, replacing it with an average value, and adding a partial missing feature value A binary missing indication feature value; the text note obstacle analysis unit 1322 is a verbatim note content according to the text description of each of the actual pending data and the feature value to be processed, using an automatic word breaking tool and Luo Gis regression analysis, first calculate the obstacle-related word frequency combination of the text description is the probability of belonging to the client or the ISP side obstacle, and take the probability as one of the characteristic values of the subsequent model input; the high-dimensional feature value multiple combination establishing unit 1323, One or a plurality of sets of high-dimensional feature values are generated for each event to be processed for the actual estimated data and the feature value; The obstacle point classification estimation output unit 1324 uses the best optimized GBDT model parameters after the monthly update training to calculate the probability value of the obstacle area probability for each newly added obstacle period.

Please refer to FIG. 3 , which is an architecture diagram of a maintenance and operation information output module of an obstacle location system and a maintenance method for the video stream service of the present invention. The maintenance operation information output module 140 is used to evaluate different input feature values. The collection, the estimated error rate of the obstacle region estimation model after the optimization in the training phase, the output integration information and the maintenance priority scheme are used by the maintenance and management personnel respectively.

The maintenance operation information output module 140 further includes a model performance indicator establishing unit 141, which is a reference calculation model for evaluating the quality of the prediction model by using a weighted error rate evaluation method; the model error rate calculation unit 142 is Using the weighting error rate calculation formula of the model performance indicator establishing unit, the error rate of the estimated obstacle area corresponding to the model corresponding to each type of feature value set after the optimization of the machine learning training stage is calculated; the maintenance information output unit 143 is Integrate the output of the customer data to be processed, the obstacle interval judgment result and the reference model error rate to provide the maintenance personnel to select and use the corresponding inspection recommendation information according to the priority of timeliness or correctness; the maintenance priority scheme output unit 144 In order to make the maintenance priority decision scheme output of the client and the ISP in the organization based on the statistical average of the obstacle interval from the current to the past month, the priority in the recent fault zone can be prioritized. Handling and repairing, in order to achieve the maximum benefit of the use of maintenance resources.

Please refer to FIG. 4 , which is a flowchart of an obstacle positioning system and a maintenance method for the video streaming service according to the present invention. The method includes: Step 1 : The S310 data source module extracts each type of high dimension by the feature value extraction module. Estimated eigenvalues; Step 2: S320 through machine learning training and implementation module processing, first training the optimized prediction model with training data, and providing the eigenvalues to be tested according to the actual customer complaint case in subsequent implementation The data estimates the probability of the corresponding client and ISP terminal under multiple types of eigenvalues. Step 3: S330 is finally responsible for generating the operation mode selection and priority decision information by the maintenance operation information output module; Step 2 S320 For the process of machine learning training and implementation module processing, please refer to Figure 5, firstly, according to the execution time, it is judged whether the model training needs to be re-executed. For example, the one-month cycle is used as the operation interval, and the operation is re-trained every other month. The unit performs a model training (steps 1 to 5 below). If the model training is not required during normal operation during the month, it is carried out by the implementation unit. According to the procedure, the steps are as follows (step one, step six to step nine), including: step one, S410 whether the training model is generated, if yes, the first simulation training is performed first, and the S420 training target is established, if not, then Perform S430 category and missing value pre-processing; Step 2: After the S420 training target is established, enter S421 category and missing value pre-processing training; Step 3, S422 text note obstacle analysis training; Step 4, S423 high-dimensional feature value multiple combination Establish training; Step 5, S424 optimization obstacle classification model establishment, and return to S430 category and missing value pre-processing; Step six, S431 text note obstacle analysis; Step seven, S432 high-dimensional feature value multiple combination establishment; Step eight , S433 obstacle point classification estimated output; step IX, determine whether S440 calculates the next user, if yes, then return to S410 whether to generate a training model, if not, then S450 ends.

It can be known from the above steps that the S420 training target is established in step 1. In order to establish the obstacle point determination model, we must first provide the accuracy rate judgment target of the machine learning algorithm model training. For example, we use all the video stream service obstacle processing history records in the last month to complete the repair as the training target. The practice is to mark the obstacle area as two types according to the obstacle reason and repair report data finally reported by the maintenance personnel. It is the client barrier that is closer to the user side of the obstacle point, and the ISP side barrier that is closer to the aggregation and core network, and the problem of the obstacle point determination is transformed into a binary classification problem. In the history of obstacle processing history in this month, according to time, the first 70% were taken out as model training and optimization data, and the last 30% were used as follow-up model performance index test data.

Step 2 of the S421 category and the missing value pre-processing training. For the collected feature values, if there are non-numeric categorical eigenvalues, we expand them into binary indicator eigenvalues (for example, can be (1) , 0) means male and (0, 1) means female. If the numerical characteristic value has a missing value, it is replaced by the average value, and a binary missing indication characteristic value is added for the partially missing feature value.

Step 3 of the S422 text note barrier analysis training, in view of the description of the text note record as a free-form text string, can not be directly used. In this regard, we propose a pre-processing method that converts a text note into a real value, representing its relevance to the ISP or client. First, use the Chinese word-breaking tool, such as the Jieba application software, to break words in the Chinese description of each text note and the reason code, and express the result in word frequency, for example (Internet access barrier, 2) or (remote control failure, 3) . Then the word frequency after the word break is used as the feature vector, and the linear Logistic Regression classification model is used for training. The trained model can estimate the obstacle point for the written note after the word break belongs to the ISP or client. The probability of the obstacle, and this probability estimate will be used as one of the input eigenvalues of the subsequent obstacle point prediction model. In order to avoid overfitting, this training process also collects independent training materials for the analysis of the text notes.

Step 4: S423 high-dimensional eigenvalue multiple combination establishment training, in the establishment of training model with high-dimensional feature value set, high dimension refers to all feature parameters, including numerical parameters and category parameters, fully expanded to include more than 300 features vector. In addition, in order to meet the maintenance needs of different characteristics, it is divided into a variety of combinations:

Type 1 eigenvalue input set: including service quality management system feature value extraction unit, obstacle report management system feature value extraction unit, loop quality diagnosis management system feature value extraction unit, and broadband network monitoring system feature value extraction unit, four All the feature values extracted by the extraction unit, the feature value set of the model is complete and the accuracy is high, but the feature value of the manual diagnosis pre-test code in the feature value extraction unit of the obstacle notification management system needs to be sent Work orders can only be obtained after performing manual tests, so the timeliness is low.

The eigenvalue input set of the type 2: as the eigenvalues produced by the units of the type one, only the manual diagnostic pre-test code in the feature value extraction unit of the obstacle management system needs to be removed. The eigenvalue set of this type has higher timeliness because it does not contain the manual diagnostic test part, but the accuracy of judgment can be improved by manual diagnosis, so the accuracy of type one is higher than that of type two.

Step 5: S424 optimizes the obstacle point classification model, establishes a usable high-dimensional eigenvalue classification model, and uses the Gradient Boosting Decision Tree (GBDT) as the selected prediction model, and uses high-dimensional eigenvalues. Multiple combinations establish multi-type eigenvalue input sets in the training unit, and train their respective optimal optimization models.

Gradient lifting decision trees are common classification algorithms in the field of machine learning. Compared with many common machine learning methods, the gradient decision tree has the advantages of no feature scaling and a non-linear feature combination. This module uses the gradient boosting method to build a decision tree in order, optimizes the defined loss function, and finally outputs the best decision trees established.

After multiple types of eigenvalue input of training output are established by multi-dimensional combination of high-dimensional eigenvalues and training is completed, GBDT algorithm outputs T decision tree, and its prediction function is defined as f _T , and a new customer application case is given in the future implementation. For all relevant feature value set data, let its feature vector be x _test , we can use the following formula (1) to evaluate the obstacle point of the customer's application case as ISP end (y=1) or client obstacle (y=-1) The chance.

In the implementation unit, the S430 category and the missing value pre-processing, the S431 text note obstacle analysis, and the S432 high-dimensional eigenvalue multi-combination are established in step 5 of the implementation unit, and the S421 category and the missing value pre-processing training in the training unit respectively. S422 text note obstacle analysis training, and S423 high-dimensional eigenvalue multi-combination establishment training is the same, the difference is only processed in the implementation unit is not the training data but the actual pending processing calculation of the feature value data.

After the implementation unit performs the pre-processing of category and missing value, the analysis of the text note barrier and the multi-dimensional eigenvalue multi-combination, the GBDT prediction model can be used to predict the output of the obstacle classification. The prediction function is defined as f _T From the actual case type of the characteristic value set data vector x _test , again using the above formula (1) to evaluate and produce the probability that the obstacle point is the ISP end (y = 1) or the client side obstacle (y = -1).

The process of the operation mode of the third step S330 is as shown in Figure 6. After the estimated model calculates the predicted value of the obstacle area, the maintenance operation information output module will be responsible for outputting the final integrated maintenance information and Maintain priority decision information, including: Step 1, S510 design model performance index evaluation, no change after the first design is completed; Step 2, S520 according to the latest training phase of the reserved test data, calculate the model error rate; Step 3 S530 outputs the information to be provided to the maintenance personnel; Step 4: S540 outputs the maintenance priority plan information to be provided to the management personnel; Step 5: S550 determines whether there is a next user to be estimated, if Yes, return to step 2, S520 calculates the model error rate according to the reserved test data of the latest training phase, and if not, ends.

The evaluation of the performance index of the S510 design model in step 1 is no longer changed after the first design is completed. It is the obstacle area estimation result of the optimization model of the design model performance index evaluation unit to evaluate the various characteristic parameters into the machine learning training stage. The error rate indicates the quantified value of the model performance. It is not changed after the first design. In this case, the model performance is verified by a weighted error rate. The model performance weighted error rate formula is as follows.

Err is the weighted error rate, w _i is the weight representing the severity of the estimated error. According to the maintenance difficulty experience, we set the client barrier weight to 1 and the ISP barrier weight to 1 to the value. The level of impact of the obstacles is large, so it is estimated that the mistake will cause a large loss. p _i ≠y _i For the indication function, if the model prediction category p _{i is} not equal to the actual obstacle point y _i , its value is 1 and vice versa.

For example, in a training test data, there are 4 obstacle cases. The fault area is sequentially {client, client, ISP, client}. After the model is estimated, the fault area is estimated as {client, client, ISP. , ISP side}. In the last obstacle case in this case, the prediction area is wrong. The unweighted error rate is 25%. If the client weight = 1 and the ISP weight = 5, the weighted error rate is: Err=(0+0+ 0+5)/(1+1+1+5)=62.5% formula (3)

In S520 of step 2, the model error rate is calculated according to the reserved test data of the latest training stage, and the model error rate calculation unit function is to calculate the weighted error rate calculation formula (2) established by using the above model performance index, and calculate the completion of the training phase. The error rate of the estimated model corresponding to each optimized set of feature values. The historical report data of obstacle repair has been completed in the previous month, sorted by time, the first 70% is used as model training and optimization data, and the last 30% is retained as the model error rate calculation unit to calculate the estimated model error rate. Use. Subsequent monthly performance evaluation values are calculated in the same way. According to the test data of 30% model performance index reserved in the last training phase, the prediction model error rate of a plurality of different input feature value sets is calculated by the formula of the model performance index establishing unit, and the output result examples are as follows:

The eigenvalue input set of the type one: the error rate is 2.13%.

The eigenvalue input set of the type 2: the error rate is 2.26%.

Among them, type one and type two eigenvalues, such as high-dimensional eigenvalues, are combined in the training to establish the training. The former emphasizes correctness and the latter emphasizes timeliness.

In S3 of step 3, the output information to be provided to the maintenance personnel is used to output the integrated transportation information used by the maintenance personnel, including the basic information and the estimated information of the threshold number of the pending obstacle area. The obstacle area, and the error probability of the various types calculated by the model error rate calculation unit, the output result examples can be as follows:

The difference between Type 1 and Type 2 is that the time required to produce the estimated value and the input eigenvalue are different. For example, the high-dimensional eigenvalue multi-combination establishment training, the former focuses on correctness, and the latter focuses on timeliness. Under normal circumstances, the maintenance unit may select the first type of estimated area with higher correctness for maintenance. If very fast maintenance is required, for example, the customer is a national defense or important customer of the people's livelihood or has a strict SLA (Service Level Agreement) contract. Then, after the type one estimated value has not been calculated, the selectable type 2 suggestion area is quickly advanced to the processing.

The S540 of step 4 outputs the maintenance priority plan information to be provided to the manager, and outputs the decision information of the maintenance priority plan used by the manager, including the recommended quantitative value of the maintenance priority of the client and the ISP. For example, an ISP company has two sets of maintenance personnel and equipment for maintaining client and ISP equipment in a certain six service areas. First, the client is used as an example. The clients in the last 30 days and the last month are represented by A and B respectively. The average number of obstacles is estimated, and the threshold for evaluating whether the ratio of A to B is too large is set to T (T>0 and T<1). If T>|(AB)/B|, it means that the absolute value of the average of the last 30 days and the average value of the previous month is less than T, then the maintenance priority is represented by 0 if the number of such pieces does not change much; |(AB)/B|>=T, that is, the absolute difference between the two is above T (inclusive), then the maintenance priority is indicated by 1 when the obstacle increases in the last 30 days (ie A>B), and when the obstacle is reduced (ie A<B) Maintenance priority is indicated by -1. The maintenance priority value of the ISP is calculated in the same way as the client. However, the threshold value T can be set according to the model performance weighted error rate weight of the formula (2), which is set to be different from the client. For example, when the ISP-side weighted error rate weight is 5 times that of the client, it means that the ISP end is more important, and the ISP-side gap threshold T can be set to 1/5 of the client. The total maintenance priority of the district directly adds the maintenance priority values of each region. Higher values indicate that more processing priorities are needed in the near future, and it is suitable for deploying higher-level maintenance equipment and more experienced maintenance personnel, or from maintenance areas and equipment. An example of the output of the output unit of this maintenance priority plan can be as follows:

This example shows that the company recently took the Taichung service area as an example. It suggested reducing the priority of client maintenance and increasing the priority of ISP maintenance. The Hsinchu service area can maintain the status quo. In general, the whole area must add clients. Maintain priority and reduce ISP end maintenance priority, or transfer too many maintenance equipment and senior maintenance personnel at the ISP end to the client.

The detailed description of the present invention is intended to be illustrative of a preferred embodiment of the invention, and is not intended to limit the scope of the invention. The patent scope of this case.

To sum up, this case is not only innovative in terms of technical thinking, but also has many of the above-mentioned functions that are not in the traditional methods of the past. It has fully complied with the statutory invention patent requirements of novelty and progressiveness, and applied for it according to law. Approved this invention patent application, in order to invent invention, to the sense of virtue.

Claims (13)

 An obstacle positioning system for video streaming services includes: a data source module, which is a complex information source for collecting and determining obstacle positioning; and a feature value extraction module, which is based on various types of information required for positioning obstacles. Different source system characteristics are extracted to form an input feature parameter group of the subsequent machine learning analysis module, and the feature values of each unit in the data source module are processed; the machine learning training and the implementation module are accepted The eigenvalues of the eigenvalue extraction module are output, and after further data pre-processing, the machine learning is used to train and obtain the optimized model parameters, and further includes a training unit and a real-life unit; the maintenance operation information output module In order to establish an evaluation method that can simultaneously consider the weight of the estimated error severity of the client and ISP-side obstacle areas, evaluate different input feature value sets, and optimize the obstacle area prediction model after the machine learning module training stage is completed. The error rate, the output is suitable for the integration of different user characteristics, and the dimension based on the quantitative value of maintenance priority Priority programs are for maintenance and operation and management personnel.    The image locating service obstacle locating system according to claim 1, wherein the data source module further comprises: a service quality management unit, which includes an image quality level and service quality data of the application layer; The management unit is related to the obstacle report related information of the image service; the loop quality diagnosis management unit includes the loop quality test data of the physical layer and the special type user loop construction method record; the broadband network monitoring unit includes the nodes of the ISP industry. Equipment brand model data, equipment alarm code and alarm code and content information.    The obstacle location system for video stream service according to claim 1, wherein the feature value extraction module further comprises: a service quality management feature value extraction unit, which is to obtain a high quality user of 4K or higher. Binary flag value, recent video stream service quality indicator, video stream service appeal probability; obstacle report management feature value extraction unit, to obtain video stream service report reason code, report description text note, manual diagnosis pre-test Code; loop quality diagnostic management feature value extraction unit, to obtain digital Subscriber Line Access Multiplexer (DSLAM) brand model, DSLAM firmware version, voice and audio band attenuation value, uplink SNR (Signal To Noise Ratio), downlink SNR, periodic quality monitoring value of the client and ISP, whether to use the special method binary flag value; broadband network monitoring feature value extraction unit to obtain the video streamer box or home The model of the multi-function gateway, the downstream rate of the video streamer box or the home multi-function gateway, and the terminal design Type, quantization index values central office equipment alarm, alarm type term frequency (term frequency), alarm severity index value.    For example, the obstacle positioning system for the video streaming service according to the first aspect of the patent application, wherein the training unit of the machine learning training and the implementation module further comprises: a training target establishing unit, which is a judgment for establishing an estimation model. As a benchmark for calculating the loss function and optimization during the training model; the class and missing value pre-processing training unit preprocesses the eigenvalues of the training data and includes expanding the categorical eigenvalues into binary indication features. Binary indicator, and if the numerical eigenvalue has a missing value, it is replaced by the average value, and a binary missing indication eigenvalue is added for the partially missing eigenvalue; the text note obstacle analysis training unit is For each incident to be processed in the training data, according to the textual description of the verbatim notes, using the automatic word-breaking tool and Logis regression analysis, first calculate the barrier-related word frequency combination of the text description belongs to the client or ISP. The probability of the obstacle, and the probability is taken as one of the characteristic values of the subsequent model input; the high dimensional eigenvalue multiple The training unit is established to make one or a plurality of high-dimensional feature value sets for each obstacle to be processed in the training data; the optimization obstacle point classification model building unit is to use a nonlinear gradient lifting decision tree (Gradient Boosting Decision) Tree, GBDT) is the main estimation model. After inputting the multi-dimensional combination of high-dimensional eigenvalues to establish the combination of various types of high-dimensional eigenvalues produced by the training unit, the optimal model parameters of the training data are found through the optimization process of minimizing the loss function. For each application, it is estimated that each new interval of the obstacle to be judged.    The obstacle location system for the video streaming service as described in claim 1, wherein the machine learning training and implementation module implementation unit is responsible for actual real-time data processing estimation, and includes: categories and omissions The pre-processing unit preprocesses the eigenvalues of the actual data to be estimated and the eigenvalues, and includes expanding the categorical eigenvalues into binary indicator eigenvalues, and if the eigenvalues are numerical eigenvalues If there is a missing value, it is replaced by the average value, and a binary missing indication feature value is added for some missing feature values; the text note obstacle analysis unit is for each obstacle in the actual data to be estimated and the feature value to be processed. The event, according to the textual description of the verbatim note content, using the automatic word-breaking tool and the Logis regression analysis, first calculate the obstacle-related word frequency combination of the text description is the probability of belonging to the client or the ISP side obstacle, and the probability And as one of the characteristic values of the subsequent model input; the multi-dimensional eigenvalue multi-combination establishing unit is for the actual data to be estimated and the eigenvalue An obstacle waiting to be processed, making one or a plurality of high-dimensional feature value sets; the obstacle point classification estimating output unit, using the best optimized GBDT model parameters after monthly update training, calculating each new pending obstacle interval Judgment value of the probability of the obstacle area of the case.    The obstacle positioning system for the video streaming service according to claim 1, wherein the maintenance operation information output module further comprises: a model performance indicator establishing unit, which establishes an evaluation by a weighted error rate evaluation. The benchmark calculation model for predicting the quality of the model; the model error rate calculation unit is to calculate the weighted error rate calculation formula of the model performance indicator unit, and calculate the corresponding estimate of each type of feature value set after the machine learning training phase is optimized. Model error rate; the maintenance information output unit is to integrate the output of the customer data to be processed, the obstacle interval judgment result and the reference model error rate, so as to provide the priority of the maintenance personnel according to the timeliness or correctness. Maintenance recommendation information; maintenance priority program output unit, based on the last 30 days and the past month within the barrier interval statistical average, the organization of the client and ISP end of the maintenance priority decision-making program Out, so that more fault areas can be prioritized and repaired in order to achieve maintenance resources The biggest benefit.    The obstacle positioning system for the video streaming service according to the second aspect of the patent application, wherein the service quality data of the application layer is an image quality level and a numerical quality indicator of each image terminal device.    For example, the obstacle positioning system for the video streaming service described in claim 2, wherein the related information of the obstacle is the content of the application, the description text and the test code after the artificial pre-test.    For example, the obstacle positioning system for the video streaming service described in claim 2, wherein the loop quality test data is a line coupling (bundling) and a light copper hybrid (G.fast) special method record, and an electrical characteristic estimation of the line. The value and the periodic record of the switching office equipment of the ISP closest to the user end.    For example, the obstacle positioning system for the video streaming service described in claim 3, wherein the manual diagnosis pre-test code is a reason code for the obstacle input after the preliminary diagnosis by the professional diagnostic personnel.    A method for locating and locating an image stream service includes: step 1: the data source module extracts the high-dimensional types of estimated feature values by the feature value extraction module; and the second step, through machine learning training and real For module processing, the optimized prediction model is trained with the training data, and the corresponding client and ISP terminal under multi-type eigenvalues are estimated based on the eigenvalue data of the actual customer complaint case. Estimate the probability of obstacles; Step 3: Finally, the maintenance operation information output module is responsible for generating the operation mode selection and priority decision information;    The method for locating and locating an image streaming service according to claim 11 wherein the process of the machine learning training and the implementation module includes: step one: whether a training model is generated, and if yes, first Perform the first simulation training, establish the training target, if it is no, perform the category and the missing value pre-processing; Step 2, when the training target is established, enter the category and the missing value pre-processing training; Step 3, the text note obstacle analysis Training; Step 4, multi-dimensional combination of high-dimensional eigenvalues to establish training; Step 5, optimization of obstacle classification model establishment, and return to category and missing value pre-processing; Step 6, text note obstacle analysis; Step VII, high-dimensional features Multiple combinations of values are established; Step 8: Obstacle point classification predicts output; Step 9: Determine whether to calculate the next user, if yes, return to whether the training model is generated, and if it is no.    For example, the method for locating and locating the video streaming service according to claim 11 is as follows: the process of the operation mode includes: step one, evaluation of the performance index of the design model, and no change after the first design is completed; 2. Calculate the model error rate according to the reserved test data of the latest training phase; Step 3: The output shall be provided to the maintenance personnel for the maintenance information; Step 4: The output shall be provided to the management personnel for the maintenance priority program information. Step 5: Determine whether there is a next user to be estimated. If yes, return to step 2, and calculate the model error rate according to the reserved test data of the latest training phase. If not, the process ends.